OpenAI Releases 722 Manuscripts of Mathematical Results

OpenAI has released hundreds of mathematical results produced by an unpublished model, while detailing its dissemination choices and estimated costs per result. An independent committee, AGMAI, is calling for regulations around these announcements and to prevent their marketing use, as the real impact will need to be assessed by the community.

Safeguards Established by AGMAI and an Already Lively Debate

At the end of September, AGMAI, an independent advisory group composed of elite mathematicians, published its first recommendations. It encourages AI laboratories to quickly publish their results through established academic channels when possible, and to provide information such as the name of the model used, the prompts, and the computational costs. AGMAI also requests that AI companies refrain from using the publication of mathematical results for promotional purposes, deeming this practice detrimental to the mathematical community. Meanwhile, the speed at which AI laboratories publish their advancements fuels a debate on research practices, ethics, and how the work of human mathematicians, used as a basis for AI systems, is credited.

What the Release Contains and How It Is Disseminated

OpenAI has published 722 manuscripts, divided into 372 sets of results. The files available online provide summaries of the model's reasoning, assessments of the computational resources used, as well as quantitative data on the number of problems addressed. The results are accessible via a GitHub repository, accompanied by protocols for review and citation, and OpenAI specifies that it continues to explore other community hosting solutions to comply with the committee's recommendations. For its upcoming publications, the company plans to improve the quality of the articles by working on citations, mathematical formalism, and the presentation of results. According to OpenAI, each result required an average of three hours of thought from ChatGPT Pro. The solutions come from a cutting-edge unpublished model.

Scope, Precedence, and Uncertainties About Impact

OpenAI has disclosed solutions to several long-standing unsolved mathematical problems, while AGMAI mentions answers to "hundreds" of open questions. These advancements had been anticipated for several weeks, with no information provided regarding the problems addressed or the publication date. In September, OpenAI stated that its model had solved more than 100 open problems across most areas of mathematics. The full impact of this release will take time to be assessed by the mathematical community. These new results add to an expanding corpus, including works from OpenAI and competing laboratories like Anthropic, with advancements reaching up to a Millennium Prize problem. AGMAI intervenes in this context where breakthroughs and ethical questions coexist.